Repository navigation
v0.1.8 — hardening & bug fixes
[0.1.8] - 2026-07-07
3-layer debug lands (app, HCI, and radio), alongside a hardening pass for the CRA Readiness Check driven by real field runs on Windows and macOS.
3-layer debug
Adsum now debugs across all three layers of a BLE connection: the application log, the HCI host↔controller bus, and the over-the-air radio, correlated to show where a flow actually broke. It builds on the HCI decoding shipped in 0.1.7 (parsing Host Controller Interface event streams into human-readable BLE protocol events). The guided HCI + Sniffer sample run promised in 0.1.7 is now the on-ramp: a walkthrough of a real one-directional BLE bug that lands on the one-line fix and bridges into the CRA readiness check. It runs on a bundled sample with no hardware needed, and the same debugging works on your own nRF boards.
CRA runs don't dead-end anymore
- A CRA run now rests on an open question, never a "task complete" box. No more dead-end endings or "I'll continue later" traps, the run always offers concrete forward actions (triage this CVE, start closing this gap), and you leave simply by moving on. The old completion scorecards (with pass/fail glyphs) are blocked at the source.
- Knowledge-loading is self-healing. A transient network blip on a knowledge fetch now retries silently; a mistyped knowledge path auto-corrects when the catalog has exactly one match (a real run dead-ended on
cra/rules/core.mdvscra/core.md, that class of failure is gone). Error messages now say precisely what failed: transient fetch vs not-in-catalog vs registry unreachable. - The readiness-report integrity guard is fairer and clearer. It no longer misreads honest phrasing like "62 total packages: 10 queryable" as a wrong count (a correct report was rejected 3× for this), and every rejection now quotes the exact line it objected to, so a rewrite lands in one attempt.
Honesty & safety hardening (from real run reviews)
- Never weakens your project to make a scan work. New hard rules: the agent must never disable your security features (secure boot, flash encryption, signed OTA) to force a build, must never edit your SDK/toolchain installation (a run had patched a script inside
C:\ncs\, now banned), and if an earlier scan run left your config modified, the posture check detects it and offers a restore instead of counting those as your gaps. - ESP SBOM generation fixed for IDF 5.x. The documented
--output-fileflag is used, andidf.py sbom-create --spdx-file(absent on IDF 5.5.4) is version-checked before use instead of failing. - Silent commands aren't "failures" anymore. Commands that legitimately produce no output (
mkdir,cp, …) no longer report a scary "technical issue", the agent is told plainly: silent success, verify state directly if it matters.
Windows reliability
- The terminal just works on a fresh Windows install. New Windows machines ship PowerShell locked down (Restricted execution policy), which silently blocks VS Code's shell integration so the agent can't read command output. The extension now detects and repairs this in the background at startup, sets the execution policy to RemoteSigned (current-user scope), selects a working default terminal profile, and restarts stale terminals, with a dismissible note of what changed. Group Policy-managed machines are left untouched.
Compact input area
- The input stack is dramatically slimmer: one-line input that grows as you type, auto-approve as a compact ⚡ chip next to @, and the wide Cancel/Resume buttons replaced by a Claude-style send ↔ stop morphing icon (brand cyan, instant tooltips). The chat input glows cyan on focus.
- The "What's new" card icon is now theme-consistent (no more OS-style emoji).